Browser Inference Has Only Two Real Hurdles: WASM Speed and WebGPU Gaps
gdechichi · x · 2026-09-25
A developer lays out why running models in the browser is closer than expected: the browser has access to the same GPU and CPU cores as the native machine, so hardware isn't the bottleneck.
The two remaining challenges are that WASM is still slower than native assembly, and WebGPU lacks some modern graphics API features. Quipped the author: if you want speed, you write C.
Related event: Benchmarks Show Browser Games Match Native; Only WASM and WebGPU Lag(3 posts)→
More from Infra
- SpaceX unveils AI training cluster site: $90B+ invested, 7,500 local jobs, 3.3 GWh Megapacks — elonmusk · 2026-09-26
- OpenAI Has ~1.9 GW of Compute, Wants 30 GW; US Buildout Nears 100 GW — le_james94 · 2026-09-26
- Inference Is the COGS of AI: Gross Margin Is Set by Token Cost and Speed — le_james94 · 2026-09-26
- One Megawatt of AI Factory Costs $59M Up Front, Pays Back in Just Over Four Years — le_james94 · 2026-09-26
- AI Stack Revenue Grew 5x to $435B in Two Years, Yet Semiconductors Take 79% of Profit — le_james94 · 2026-09-26
- Token Prices Fell 99% in 2.5 Years Yet Total Spend Rose; H100 Spot Climbed All Along — le_james94 · 2026-09-26